path integral molecular dynamics ...helper.ipam.ucla.edu/publications/ccsws4/ccsws4_9912.pdfpath...

PATH INTEGRAL MOLECULAR DYNAMICS:APPLICATIONS & NEW ALGORITHMS.

Alejandro Perez PazNano-bio Spectroscopy Group

University of the Basque CountrySpain

IPAM CCS WorkshopLos Angeles, CA

May 18, 2011

Outline

I Introduction to Path Integrals (PI)I Double proton transfer in DNA base pair modelsI Development of free energy methods in PIMDI Improving the convergence of PIMDI Summary

Outline

Introduction to Path Integrals

Double proton transfer in DNA base pair models

Development of free energy methods in PIMD

Improving the convergence of PIMD

Summary

Quantum Mechanics in Nature

What can PI do for you?

I DelocalizationI Zero Point Energy (ZPE)I TunnelingI Theory development (QCD)I quantum dynamics? R. P. Feynman

www.wikipedia.org JCP 130, 184105 (2009).

www.wikipedia.org

PI Formulation of Canonical Partition Function

The canonical density matrix operator

ρ (β) = exp (−βH) ,

where H = T + U and [T,U] 6= 0.

Z (β) = Tr(

e−βH)

=

∫dx1〈x1|e−β(T+U)|x1〉 = lim

P→∞

∫dx1〈x1|ΩP|x1〉,

where Ω = e−β2P Ue−

βP Te−

β2P U, which follows from Trotter

theorem1

e−β(T+U) = limP→∞

[e−

β2P Ue−

βP Te−

β2P U]P.

1Tuckerman, Stat. Mech.: Theory & Molecular Simulation, Oxford, 2010

Inserting (P− 1) completeness relations I =∫

dx|x〉〈x|

Z(β) = limP→∞

∫dx1dx2 · · · dxP〈x1|Ω|x2〉〈x2|Ω|x3〉 · · · 〈xP|Ω|x1〉

= limP→∞

∫ P∏i=1

dxi〈xi|Ω|xi+1〉|xP+1=x1 .

Need to compute

〈xi|Ω|xi+1〉 = 〈xi|e−β2P Ue−

βP Te−

β2P U|xi+1〉

= e−β2P

[U(xi)+U(xi+1)

]〈xi|e−

βP T |xi+1〉.

What can we do...?

Then, we obtain

〈xi|Ω|xi+1〉 =

√mP

2πβ~2 e− mP

2β~2 (xi+1−xi)2− β

2P [U(xi)+U(xi+1)].

Path integral representation of canonical quantum PF

Z (β) = limP→∞

(mP

2πβ~2

)P/2 ∫xP+1=x1

dx1 · · · dxP e−βUeff ,

where Ueff =∑P

i=1

[mP

2β2~2 (xi+1 − xi)2 + 1

P U (xi)].

Each quantum particle is mapped onto a collection of Pinteracting quasi-particles.11Chandler & Wolynes, JCP 74 4078 (1981).

The ring-polymer isomorphism

Analytical solutions to path integrals are limited:devise numerical techniques for complex systems.Discretized version of the path integral quantum PF:

Z(β) = limP→∞

(mP

2πβ~2

)P/2 ∫dx1 · · · dxP e

−β∑P

i=1

[m2 ω

2P(xi−xi+1)2+

U(xi)P

],

where ω2P = P/ (β~)2 and xP+1 = x1.

A quantum particle ismapped onto a collection ofP “beads” coupled viaharmonic springs & eachunder external potentialU/P.

centroid xc = 1P

∑Pi=1 xi

The ring-polymer isomorphism

Introducing nuclear quantum effects with the ring polymer.

www.telluridescience.org/TullyLectures/tully.pdf

www.telluridescience.org/TullyLectures/tully.pdf

PIMD Implementation

Assuming ergodicity, sampling can be effected using MD.Introduce a set of uncoupled Gaussian integrals to enablesampling of phase space (fictitious m′k)

Z(β) ≈ N∫

dp1 · · · dpP dx1 · · · dxP e−β∑P

k=1

[p2k

2m′k+ m

2 ω2P(xk−xk+1)2+ 1

P U(xk)

].

I Remember: Time is only a parameter for the exploration ofphase space (unphysical dynamics).

I Naive PIMD implementation suffers from non-ergodicity &multiple time scales issues1.

1 Hall & Berne, JCP 81 3641 (1984).

Efficient PIMD ImplementationSolution M. Tuckerman et al. JCP 99 2796 (1993).

I Transformation coordinates:normal modes uk = 1√

P

∑Pj=1 xje2πi(j−1)(k−1)/P (centroid is u1)

or staging u1 = x1, uk = xk − (k−1)xk+1+x1k , k = 2, · · · ,P.

H (u,p) =

P∑k=1

[p2

k2m′k

+mk

2ω2

Pu2k +

1P

U (xk (u))

],

where m1 = 0. mk = kk−1 m for staging and

mk = 2mP [1− cos (2π(k − 1)/P)] for normal modes.I Fictitious masses chosen to collapse all time scales

m′1 = m, m′k = mk

I Massive thermostatsAlthough computationally intensive, PIMD can be parallelizedextremely well.

Outline





Summary

Proton transfer reactions

Ubiquitous and crucial process in nature

S1 − H∗ · · · S2 ⇐⇒ S1 · · ·H∗ − S2

I Quantum Effects can be important even at roomtemperature!1

I Rearrangement of chemical bonds calls for ab initiodescription of the electronic part

1 Tuckerman & Marx, PRL 86 4946 (2001).

Intermolecular proton transferDNA base pairs

I “Rare” tautomers implicated in DNA mutationsI Proton tunneling was hypothesized to favor “rare”

tautomers1

I Experimental evidence only for analogues (large KIE)2

1 Lowdin, RMP 35 724 (1963). 2 Zewail et al., Nature 378 260 (1995).

Experiment on 7-azaindole dimer(g)

Douhal, Kim, & Zewail et al., Nature 378 260 (1995).

Rare tautomers of DNA bases

http://www.ncbi.nlm.nih.gov/

http://www.ncbi.nlm.nih.gov/

Intermolecular double proton transfer

Formamidine-Formamide (FIFA) tautomer complex1

I Model system for A-T DNA base pairI Concerted mechanism.

1 J. Leszczynski et al., JPCA 106 12103 (2002)

Intermolecular double proton transfer

Model for G-C DNA base pair

Forces from “First Principles”

Combining the best of two worlds...

I The quantumnature of nuclei isdescribed usingthe PI formalism

I The electronicpart (governinginteractionbetween nuclei)is approximatedusing DFT

Car & Parrinello Molecular Dynamics (CPMD), PRL 55 2471 (1985).

Choice of DFT XC functional

Energetics (kcal/mol) from TURBOMOLE code/TZVPP basis set.Structures correspond ≈ to reactants (R), the transition state(TS), and products (P)

GC model AT modelmethod ∆EP−R ∆ETS−P ∆ETS−R ∆EP−R ∆ETS−P ∆ETS−R

RHF 13.50 14.65 28.15 11.80 9.47 21.27MP2 9.37 4.04 13.41 8.62 4.37 12.99CC2 10.24 2.67 12.91 9.24 3.08 12.32PBE0 10.95 3.00 13.95 8.93 2.81 11.74PBE 10.63 0.55 11.18 8.56 1.19 9.75B3LYP 11.39 4.78 16.17 9.47 3.52 12.99BLYP 11.16 3.15 14.31 9.30 2.44 11.74

BLYP represents best compromise between accuracy and cost!

Free energy methods & rare events

I Reaction coordinate: ξ(r) = dHN − dHO

I Umbrella sampling1 and WHAM technique2

I Drive system along reaction coordinate using harmonicbias potential (centroid)3 ξc = 1

P∑P

i=1 ξi

1 Torrie & Valleau, Chem. Phys. Lett. 28 578 (1974)2 Ferrendberg & Swendsen, Phys. Rev. Lett. 61 2635 (1988)3 Gillan, PRL 58 563 (1987). Voth, JCP 97 8365 (1993).

Free Energy Profiles Double Proton Transfer

Nuclear quantum effects have a dramatic impact on free energyprofiles.

Rare tautomers predicted dynamically unstable:therefore, not involved in mutations!

Free Energy Profiles Double Proton Transfer

J. AM. CHEM. SOC. 132, 11510 (2010).

www.youtube.com

www.youtube.com

Outline





Summary

Computing Thermodynamic Quantities: Estimators

An estimator: function whose average approximates a physicalobservable.

Example: E(β) = − ∂∂β ln Z(β)

〈E〉 ≈ 1ZP

(mP

2πβ~2

)P/2 ∫dx1 · · · dxP εprim e−β

∑Pk=1

[m2 ω

2P(xk−xk+1)2+

U(xk)P

],

where the primitive estimator for energy is

εprim =P

2β+

P∑k=1

[U(xk)

P− m

2ω2

P(xk − xk+1)2]

xP+1=x1

.

Use MC, MD to sample this configuration integral.A more efficient estimator (virial) was proposed:Herman et al JCP 76 5150 (1982).

Computing Thermodynamic Quantities: Estimators

An estimator: function whose average approximates a physicalobservable. Example: E(β) = − ∂

∂β ln Z(β)

〈E〉 ≈ 1ZP

(mP

2πβ~2

)P/2 ∫dx1 · · · dxP εprim e−β

∑Pk=1

[m2 ω

2P(xk−xk+1)2+

U(xk)P

],

where the primitive estimator for energy is

εprim =P

2β+

P∑k=1

[U(xk)

P− m

2ω2

P(xk − xk+1)2]

xP+1=x1

.

Use MC, MD to sample this configuration integral.A more efficient estimator (virial) was proposed:Herman et al JCP 76 5150 (1982).

Internal Energy: Virial vs Primitive EstimatorsThere is a more efficient estimator for the internal energy calledthe “virial”. Standard Deviation: 0.522(prim); 0.220(vir).

εvir =1

2P

P∑k=1

xk∂V/∂xk +1P

P∑k=1

V (xk) ,

Instantaneous values of virial (orange) and primitive (black)estimators for V (x) = x2/2 at β = 10,P = 64.

0 2e+05 4e+05 6e+05 8e+05 1e+06Time Step (dt=0.01)

-1

-0.5

0

0.5

1

1.5

Tot

al E

nerg

y (n

atur

al u

nits

)

PrimVir

Isotope Effect via Thermodynamic Integration (TI)

Motivation: No PI method to study Isotope Effect (IE).Enzyme catalysis: proton transfer is the rate limiting step in

many enzymatic reactions (dehydrogenases): large IE.Acid dissociation, etc.

Typically, Site–H reacts faster than Site–D.

∆F = F(mf)− F (mi) =

∫ 1

0dλ(

dFdλ

),

where F = − 1β ln Z is the (quantum) free energy.

For a linear path m (λ) = λmf + (1− λ) mi.

Isotope Effect via Thermodynamic Integration (TI)

Z(β) ≈n∏

i=1

(miP

2πβ~2

)3P/2 ∫dnPr e−β

∑ni=1∑P

s=1

[mi2 ω

2P(ri,s−ri,s+1)

2+

U(rs)P

].

Two PI estimators for dFdλ :

primitive(dFdλ

)prim

= −N∑

I=1

(m′ImI

)[3P2β−

⟨mI

2ω2

P

P∑s=1

(rI,s − rI,s+1)2

⟩λ

],

and the virial(dFdλ

)vir

= −N∑

I=1

(m′ImI

)[3

2β+

⟨1

2P

P∑s=1

rI,s · ∂rI V (rs)

⟩λ

].

Isotope Effect via PIMD

Results for an analytically solvable system (harmonic potential)

0 0.2 0.4 0.6 0.8 1λ

-0.25

-0.2

-0.15

-0.1

dF/d

λ

exactvirprim

0 0.2 0.4 0.6 0.8 1λ

-0.3

-0.2

-0.1

0

dF/d

λ

exact

vir

prim

First derivative of quantum free energy with respect to mass.Left: m (λ) = λmf + (1− λ) mi.Right: m (λ) = λ4mf +

(1− λ4

)mi

Black: exact result. Red: virial and Blue: primitive estimators.

Isotope Effect via TI-PIMDTwo-oscillator model for O− H(D) · · ·O. (Antisymmetric Mode)Useful to compute IE & understand Ubbelohde effect.

0 0.2 0.4 0.6 0.8 1λ

-0.5

-0.4

-0.3

-0.2

-0.1

0dF

/dλ

exactprimvir

100,000 PIMD steps/window: prim=-0.2035, vir=-0.2052,exact=-0.2042.

Simultaneous change in mass and potentialChange of identity/interaction potential: Useful to study Hbinding to active sites.(

dFdλ

)=

~4√

(κ/m)coth

(β~2

√κ/m

)[mκ′ − m′κ

m2

].

0 0.2 0.4 0.6 0.8 1λ

0.1

0.2

0.3

0.4

0.5

dF/d

λ

exact

prim

vir

0 0.2 0.4 0.6 0.8 1λ

0

0.1

0.2

0.3

0.4

dF/d

λ

prim

exact

vir

First derivative of the quantum free energy with respect to mass& force constant for harmonic potential.Left: Linear path. Right: quartic path.Black: primitive result. Red: exact and Blue: virial estimators.

Perturbation TheoryMotivation: PI calculations are expensive!

∆F = − 1β

[3NP

2ln(

mf

mi

)+ ln

⟨e−β(mf−mi)

ω2P

2 A⟩

i

],

where A =∑N

I=1∑P

s=1 (rI,s − rI,s+1)2.Full PT for free energy difference due to isotope transformationusing single PIMD simulation.

100000 200000 300000 400000number PIMD steps (dt=0.01)

-0.22

-0.2

-0.18

-0.16

F(m

=2.

0) -

F(m

=1.

0)

PertExact

0 100000 200000 300000 400000number PIMD steps (dt=0.01)

-0.16

-0.14

-0.12

-0.1

-0.08

-0.06

-0.04

-0.02

0

F(m

=1.

5) -

F(m

=1.

0)

PertExact

Left: mi = 1→ mf = 2. Right: mi = 1→ mf = 1.5.Exact: red horizontal line.PT works provided change is small · · · Multistage-PT.

Perturbation Theory (II)Effect of a static electric field:

∆F = − 1β

ln 〈exp [−β (qE · r)]〉E=0.

Useful to understand ferro/paraelectric hydrogen-bonded solids(KDP), Stark Effect.

0 2e+05 4e+05 6e+05 8e+05 1e+06MD step (dt=0.05)

-0.65

-0.6

-0.55

-0.5

-0.45

F(E

=1)

- F

(E=

0)

ExactPert

-2

-1.75

-1.5

-1.25

F(E

=2)

- F

(E=

0)

ExactPert

λ-dynamics PIMD

Allow λ to vary in a continuous way using single PIMDsimulation.Example: transmutation of diatomics H2 →HCl at 300 K.

0.4

0.5

0.6

0.7

0.8

0.9

1

0 0.2 0.4 0.6 0.8 1

dF/d

λ

λ

numericalexact

∆F = F (HCl)− F (H2) = −0.00332 a.u. (exact), -0.00332 a.u.(numerical).

Isotope effects in the Zundel cation

Important in hydrogen-bonded liquids & in enzymatic reactions.

CPMD Details: Ab initio PIMD (H2 → D2).BLYP/100 Ry, Cubic Box 123 A3, P = 32, T=300 K.

Numerical (PT) ≈ 4 kcal/mol.CPL 329, 36 (2000).

Goal: change of identity in PIMD simulations

Feasible in force fields-PIMD, harder in ab initio PIMD(pseudos).

Example: protonated water/ammonia complexH2O · · · H+ · · · OH2 → H2O · · · H+ · · · NH3

Towards predicting changes in free energy barriers heights?

Predict kinetic isotope effect kH/kD ≈ e−β(

∆F‡H−∆F‡

D

).

Outline





Summary

Higher-order methods to improve PIMD averages

IDEA Why not to use the classical forces to improve averages?

⟨O⟩≈ 〈O (r1) w (r1, · · · , rP)〉

〈w (r1, · · · , rP)〉.

w (r;β) weighting function constructed using classical forces.In some cases, fourth-order convergence can be attained

without additional overhead!

G. Voth et al, JCP 115, 7832 (2001).A. Perez & M. Tuckerman, JCP (in review).

Harmonic potential at β~ω = 10

Hydrogen molecule at 300K: D [1− exp (−a (x− xeq))]

Hydrogen molecule: Ab initio PIMD

Details: T=300K, CPMD BLYP/75 Ry, Cubic Box 93 A3.

High order PIMD methods

Liquid water at ambient temperature (q-SPC/Fw force field1)

P 2nd 4th4 -0.2778 0.59898 0.0400 0.391116 0.2699 0.389232 0.3688 0.405564 0.3968 0.4067

Energy values in a.u.

I Useful to converge faster averages in quantum simulationsof bulk phases.

I Very easy to implement in existing ab initio PIMD codes.I No overhead associated.

1 JCP 125, 184507 (2006).

Outline





Summary

Summary

I The Feynman’s path integral formalism was reviewed &applied to several systems

I Nuclear quantum effects are important even at roomtemperature

I “Rare” tautomers dynamically unstable & not implicated inmutations

I New estimators were derived & implemented to investigatealchemical transformations

I Higher-order algorithm to improve convergence of PIMDestimators

Acknowledgments

The beautiful town of Baiona

Thanks for your attention! Questions?

path integral molecular dynamics ...helper.ipam.ucla.edu/publications/ccsws4/ccsws4_9912.pdfpath...

Documents